Working hypothesis · Rachel McBride
← Working frames · Rowan Advisory ↗
The Intelligence-Centred Firm
A working hypothesis about how firms may change when AI does more of the coordination and people keep responsibility for judgement.
Much of a firm’s middle is coordination: gathering information, preparing papers, moving work between teams, checking progress and turning decisions into tasks.
AI can now do more of that work. My hypothesis is that firms will need to be built differently if coordination takes less time and fewer people.
The shift
In “The Nature of the Firm”, Ronald Coase argued that firms form when coordinating work inside one organisation is cheaper than buying each piece through the market. Hierarchy, management layers and meetings became part of that machinery.
AI could reduce the cost of searching, drafting, comparing, scheduling and carrying out multi-step work. If it does, the scarce work may move upward: deciding what matters, reading the situation, choosing between competing outcomes and accepting responsibility for the result.
A hypothetical example
Imagine that a competitor announces same-day delivery. In this hypothetical case, an AI system notices the announcement, gathers the relevant customer and cost data, prepares options and coordinates the chosen response.
People still decide:
- whether the signal matters;
- which trade-offs the firm will accept;
- which actions need a named person’s approval; and
- whether the result was good enough to change the firm’s usual way of working.
The work can move faster because people are not gathering every input or drafting every step. The decision remains human because the consequences remain human.
The operating design
An intelligence-centred firm needs four things:
- A clear purpose and boundaries. The system needs to know what the firm is trying to achieve, what it may do and what it must never do.
- A record the firm can check and own. Facts, decisions and outcomes need sources, dates and named owners. The Verifiable Core is my design for this.
- AI that works over the record. Models and agents can prepare options and carry out bounded work without quietly rewriting the facts underneath them.
- Human decisions at the points of consequence. The firm names what needs approval, who gives it and what evidence they must be able to inspect.
How to begin
Rebuilding a whole organisation at once creates too much risk. Start with one repeatable workflow that matters, give a small team room to rebuild it, and run the new version beside the existing one.
- Write down the outcome, the boundaries and the person accountable.
- Bring the information for that workflow into a record you can check.
- Let AI prepare and carry out the low-risk steps.
- Keep approval at the decisions with real consequences.
- Compare the two versions using quality, speed, cost and what the people doing the work learned.
- Move the workflow only when the evidence supports it, then take on the next one.
The people problem
Entry-level work has traditionally taught people how to exercise senior judgement. If AI takes the first drafts, research and routine analysis, firms have to design a new apprenticeship instead of assuming one will appear.
One answer is to have junior people help make judgement visible: record why a decision was made, turn recurring decisions into reusable methods, test the AI’s work and study the exceptions. That is promising. It is not yet a complete answer.
What remains unproven
The technical parts are moving quickly. The human, legal and economic parts will decide how far and how fast this goes.
- How much review can people genuinely provide when the system moves at machine speed?
- How do junior people build judgement when the old training work disappears?
- Who carries liability when an agent acts across several systems?
- How does a firm change its pricing and roles while the old business still has to run?
Those questions are part of the frame. A firm built around intelligence still has to be built around people.